gpt-5.3-codex works in codex exec but fails in codex mcp-server subagent with same profile/auth

Open 💬 2 comments Opened Feb 16, 2026 by MarlonKr

What version of the Codex App are you using (From “About Codex” dialog)?

26.212.1823

What subscription do you have?

pro

What platform is your computer?

Darwin 25.1.0 arm64 arm

What issue are you seeing?

gpt-5.3-codex works for normal Codex CLI runs, but fails when the same project is invoked as a subagent via codex mcp-server.

Environment

  • Codex CLI: codex-cli 0.93.0
  • Auth: codex login status => Logged in using ChatGPT
  • OS: macOS
  • account has access to gpt-5.3-codex
  • tried with and without forced_login_method

Relevant config

Global ~/.codex/config.toml:

model = "gpt-5.3-codex"
forced_login_method = "chatgpt"

[profiles.subagent-medium]
model = "gpt-5.2-codex"
forced_login_method = "chatgpt"

[mcp_servers.CODEX_PROJECTS_SUBAGENT]
command = "codex"
args = ["--profile", "subagent-medium", "mcp-server"]
cwd = "/Users/.../xyz_agent"

Project <project>/.codex/config.toml:

model = "gpt-5.3-codex"
forced_login_method = "chatgpt"

[profiles.subagent-medium]
model = "gpt-5.3-codex"
forced_login_method = "chatgpt"

What steps can reproduce the bug?

  1. Configure project + global config as above.
  2. Confirm auth:
codex login status
  1. Run normal CLI execution in that project:
codex exec --skip-git-repo-check --profile subagent-medium "reply only: OK"
  1. Invoke same project via registered MCP subagent (codex mcp-server path), profile subagent-medium.

What is the expected behavior?

Expected behavior

Subagent call should run with gpt-5.3-codex and succeed, same as codex exec.

Actual behavior

Subagent call fails with:

stream disconnected before completion: The model `gpt-5.3-codex` does not exist or you do not have access to it.

Additional information

  • Same subagent works with gpt-5.2.
  • codex exec in the same project shows model gpt-5.3-codex and starts successfully.

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗